Machine learning algorithms using administrative data can be valuable and feasible tools for more accurately identifying opioid overdose risk, according to a new study published in JAMA Network Open.
Wei-Hsuan Lo-Ciganic, Ph.D., College of Pharmacy at the University of Florida, Gainesville, along with her research team, found that machine learning algorithms performed well for risk prediction and stratification of opioid overdose — especially in identifying low-risk subgroups with minimal risk of overdose.
Lo-Ciganic told Inside Digital Health™ that machine learning algorithms outperformed the traditional approach because the algorithms take into account more complex interactions and can identify hidden relationships that traditionally go unseen.
The researchers used prescription drug and medical claims for a 5 percent random sample of Medicare beneficiaries between January 2011 and December 2015. The team identified fee-for-service adult beneficiaries without cancer who were U.S. residents and received at least one opioid prescription during the study period.
The team compiled 268 opioid overdose predictor candidates, including total and mean daily morphine milligram equivalent, cumulative and continuous duration of opioid use and total number of opioid prescriptions overall and by active ingredient.
The cohort was randomly and equally divided into training, testing and validation samples. Prediction algorithms were developed and tested for opioid overdose using five commonly used machine-learning approaches: multivariate logistic regression, least absolute shrinkage and selection operator-type regression, random forest, gradient boosting machine and deep neural network.
Prediction performance was compared with the 2019 Centers for Medicare and Medicaid Services opioid safety measures, which are meant to identify high-risk individuals and opioid use behavior in Medicare recipients.
In order to find the extent to which patients who were predicted to be high-risk exhibited higher overdose rates compared with those predicted to be low-risk, the researchers compared the C-statistic and precision-recall curves across different method from the sample using the DeLong Test.
Low-risk patients had a predicted score below the optimized threshold, medium-risk had a score between the threshold and 10th percentile and high-risk patients were at the top 10th percentile of scores.
Based on the findings, the deep neural network and gradient boosting machine performed the best, with the deep neural network having a C-statistic of 0.91 and the gradient boosting machine having a C-statistic of 0.90.
With the gradient boost machine algorithm, 77.6 percent of the sample were categorized into low-risk, while 11.4 percent were medium-risk and 11 percent were high-risk. And with the deep neural network algorithm, 76.2 percent of people were predicted to be at low-risk, and 99.99 percent of those individuals did not have an overdose.
Lo-Ciganic said that with the promising results of the study, the next step would be to develop software to be incorporated into health systems — or an electronic health record — to see if the algorithms can be applied in real-world settings to help clinicians identify high-risk individuals.